They can mix cocktails, run marathons, and fold laundry, but despite marketing claims, humanoid robots are still far from reliably handling many different tasks on demand.
The difference was clear at the Robotics Summit in Boston in late May, where glossy marketing materials told one story while the engineers building the robots described another reality.
Humanoid Robots Move From Demos to Early Home-Style Tasks
Elon Musk has promoted his Optimus prototype, and the company recently showed it jogging in short, controlled steps. Meanwhile, Figure 03, a third-generation robot from Figure AI, is capable of independently tidying and cleaning a living room.
Companies such as China’s AgiBot and Matrix Robotics also claim their robots can welcome guests, serve coffee, and act as guides—evoking comparisons to C-3PO from Star Wars.
Chris Matthieu, a RealSense startup representative, said that operators either remotely control most humanoid robots on display or restrict them to tightly defined tasks and routines.
In practice, humans still control many of them or limit them to a single, narrow function.
Neo’s Promise of a Consumer-Ready Home Robot Still Relies on Human Control
Take Neo, the robot launched with much fanfare by 1X last October. The company promoted it as “the world’s first consumer-ready humanoid robot designed to transform life at home,” but a person off to the side still guided it in practice.
Even so, progress is real, and artificial intelligence is speeding it up. William Okazaki of sensor manufacturer Renesas Electronics said AI has “extremely accelerated” development in the field.
One of the major challenges remains robotic hands. Long considered the holy grail of robotics, they are improving: modern robots can now grasp objects with a gentle touch, and some sensors are even capable of detecting contact with human skin.
A new type of AI called a VLA (vision-language-action) model drives much of this progress. It combines written instructions with real-time camera input, allowing robots to connect what they see with what they need to do.
Another approach is the “world model,” an AI trained on huge amounts of images and video so it can anticipate how real-world situations unfold—for example, predicting how an object will move when it is squeezed.
Search for Data
But a fully versatile android capable of handling a wide range of tasks remains years away.
“For general-purpose robots, it will take longer,” said Daniel Fan of Innodisk, a company that supplies components for robots.
Many humanoid robots are already operating in real environments—such as Boston Dynamics Atlas at Hyundai facilities and Hexagon AEON at BMW sites—but these are still experimental trials rather than finished commercial products.
“You don’t really know what a robot can do until you actually see it attempt the task,” said Charlie Kemp of Hello Robot, a company that develops robots for people with limited mobility.
Data Shortages Remain a Barrier to Full Robot autonomy
According to Xinrui Bi of AgiBot, fully autonomous robots operating at scale are still not feasible, largely because there is insufficient training data.
To close this gap, companies are increasingly installing cameras in a wide range of environments to capture human activity—ranging from home cooking to industrial work in textile factories in India.
The risks are greater than with chatbots like ChatGPT, since robots operate in the physical world and their errors can cause real harm.
Valentino Fagard of Japan’s XELA Robotics, which develops touch-sensing technology for robots, said that entering more social environments requires a high level of safety for anyone interacting with them.
While engineers can impose constraints—such as limiting grip strength or restricting how close robots can get to people—there is still a challenge: like chatbots, these AI systems can be inconsistent, making their behavior difficult to fully predict.
“The challenge with what you might call world models or end-to-end VLA systems is that they’re non-deterministic and essentially black boxes,” said John Black of Brain Corp, whose robots are designed for tightly defined tasks such as floor cleaning or inventory checks.
He added that these systems are still far from meeting required safety standards, since even their developers cannot fully understand why they produce specific outputs or decisions.
Read the original article on: Tech Xplore
Read more: The Eno humanoid robot is built as a versatile, all-purpose workplace machine